Objective To address the challenges of sparse ground stations, delayed response in complex forest environments, and reduced unmanned aerial vehicle (UAV) observation accuracy caused by platform perturbations, downwash airflow, and sensor lag, this study proposes a multi-sensor UAV atmospheric monitoring and data calibration systema. This framework establishes a complete technical pipeline from data acquisition to precision enhancement, fulfilling the demand for high-spatial-resolution, dynamic atmospheric monitoring in forested regions.

Method We designed and integrated a UAV platform equipped with multi-sensor payloads, a GPRS data link, and embedded data acquisition software to simultaneously collect six pollutants (PM_2.5, PM₁₀, NO₂, SO₂, CO, O₃), and temperature-humidity data. Subsequently, we implemented a three-stage processing workflow: (1) applying the JUMP-MAD algorithm for jump detection and missing-value handling to identify and repair anomalies in raw time series; (2) generating spatiotemporally consistent “quasi-ground-truth” fields via adaptive inverse distance weighting (AIDW) interpolation, leveraging high-accuracy observations from fixed forest monitoring nodes; and (3) developing a two-stage calibration framework—RADNN-RF—that combines a deep neural network with residual architecture and sample-level attention mechanisms (RADNN) with random forest regression to learn nonlinear error patterns across multiple sensors and spatiotemporal scales.

Result (1) The outlier detection method effectively identified abrupt changes while preserving data continuity and reducing noise. (2) In leave-one-out cross-validation of interpolation methods, AIDW outperformed conventional IDW, reducing MAE from 0.823 to 0.763 and better capturing the heterogeneous concentration field in forested areas. (3) The final RADNN-RF fusion model achieved average errors below 3% across all atmospheric parameters when compared to quasi-ground-truth values. For PM_2.5, the model showed excellent agreement in trend with an R² of 0.97—significantly surpassing uncalibrated raw data. Moreover, the model enabled joint modeling across pollutants and multi-scale feature extraction, effectively leveraging auxiliary variables like temperature and humidity to enhance prediction accuracy and robustness. Calibrated concentrations closely aligned with fixed-node interpolation results and demonstrated strong generalization capability.

Conclusion The proposed three-stage pipeline (anomaly detection−spatial interpolation−deep fusion) systematically mitigates UAV observation disturbances, and substantially improves the accuracy and reliability of atmospheric monitoring in forest ecosystems. This approach offers a practical and scalable technical solution for establishing high-resolution, dynamic air quality monitoring systems in forested landscapes.